Edge Detection and Template Matching Approaches for Human Ear Detection

Transcription

1 Edge and Template Matching Approaches for Human Ear K. V. Joshi G H Patel College Engineering and Technology vallabh vidyanagar, Gujarat, India N. C. Chauhan A D Patel Institute Technology New vallabh vidyanagar, Gujarat, India ABSTRACT Ear detection is a new class relatively stable biometrics which is not affected by facial expressions, cosmetics, eye glasses and aging effects. Ear detection is the first step an ear recognition system, to use ear biometrics for human identification. In this paper, we have presented two approaches to detect ear from 2D side images. One is edge detection based method and the other is template matching method. For both the methods, the correctness the detected ear is verified using support vector machine tool. For template matching method it is also verified by Euclidian distance. The purpose the paper is also to compare the results both the presented methods. The experimental results prove the effectiveness these methods. General Terms Pattern recognition, pattern matching. Keywords Ear biometrics, ear detection, ear verification, edge detection, template matching. 1. INTRODUCTION Ear is a viable new class biometrics since ears have desirable properties such as universality, uniqueness and permanence. The ear has certain advantages over other biometrics. For example, ear is rich in features, it is a stable structure which does not change with the age. It does not change its shape with facial expressions. Furthermore, the ear is larger in size compared to fingerprints and can be easily captured although sometimes it can be hidden with hair and earrings. It has fixed background. For recognition, when an image is a side image, only the ear is unique feature from which a person can be identified. Although it has certain advantages over other biometrics, it has received little attention compared to other popular biometrics such as, fingerprint etc. Human ear detection is the first task a human ear recognition system and its performance significantly affects the overall quality the system. Ear recognition is useful for person identification when an image a side is available. The number recent researches [1, 2, 3, 4] show that recognition is possible and effective for side s by detecting and recognizing components such as ears. Hence, in this paper we present and compare two methods ear detection from 2D side images. The rest the paper is organized as follows. The proposed methods for ear detection and verification are described in section 2. The implementation and verification results are shown in section 3. The comparison both the methods is shown the section 4 and Conclusion is discussed in the section EAR DETECTION METHODS 2.1 Edge Based A block diagram for this method is shown in the figure 1[5]. Side image Skin segmentation Edge detection & connected component labeling Fig 1 : Block diagram the edge detection based method As shown in block diagram figure 1, first skin segmentation is performed from input image then nose tip is detected. After that sector containing ear is extracted. Then edge detection and connected component labeling is applied. The edges ear have been found in extracted region. From the result connected component labeling, where maximum connected edges are found a rectangle is drawn around it, that shows the detected ear Skin Segmentation Get nose point Verification the detected ear Extract sector containing ear Ear position Color Space Selection Here our input image is 2D side color image. From side image skin portion is separated out using the method suggested in [6, 7]. The goal is to remove the maximum number non- pixels from the images in order to narrow the focus to the remaining predominantly skin-colored regions. For this purpose we need to select appropriate color space from the wide variety choices such as RGB, HSV, CMYK, YCbCr etc. 50

2 From these, RGB (red-green-blue), HSV (hue-saturation-value) and YCbCr are widely used [8]. In the RGB model, each the three components may exhibit substantial variation under different lighting environments. The results YCbCr and HSV are more robust to lighting variations because in both the color spaces, color classification is done using only pixel chrominance. It is expected that skin segmentation may become more robust to lighting variations if pixel luminance is discarded and it is also verified by results. Here HSV color space is preferred for color classification because its similarities to the way human tends to perceive color. It decouples the chrominance information from the luminance information. Thus we can only focus on the hue and the saturation component Setting Threshold and Binary Image Creation After choosing the suitable color space, the next step is to separate the skin colored region from the given input image. For this, the best technique is to apply threshold. To find the appropriate values for threshold, the many images in HSV color space are examined and found out some specific ranges the components for skin color. When these threshold values are applied to the input image, the new binary image is formed in which the portions satisfying the conditions is made white and the remaining portion is made black. Figure 2 shows result skin segmentation. This is a binary image created from a RGB input image. The thresholding is done on the basis the HSV values Nose Tip Nose tip is detected by taking first white pixel from segmented skin image. The skin segmentation might not perfect and so any first non-black pixel in the binary image might not always represent a skin. So in order to obtain the nose point, the first non-black pixel is noted as a skin point only if it is surrounded by non-black pixels as well. In this way the first non black pixel in each column is noted. (a) (b) (a) Input Image (b) Segmented skin Fig 2 : Skin segmentation Once a vector the pixel locations is available, the minimum position is noted. Again, the surrounding pixels are examined in order to ensure that the identified point is a skin pixel. Figure 3 shows the identified nose point Extraction Region Face Containing Ear Once the nose point is identified, a rectangular region is extracted from the with dimensions 10cm * 20cm. The probability finding the ear in this region is high. Once this region is extracted, the ear can be located by looking for strong edges in the region. Figure 4 shows the extracted region. Fig 3 : Nose tip detection Fig 4 : Region Extraction Edge and Connected Component Labeling In this phase, edges are detected from input side image. First all color image is converted to gray scale image. For edge detection Sobel, Robert, Prewitt, Canny, Laplacian Gaussian, Zero cross etc methods are available [9]. The Canny method finds edges by looking for local maxima the gradient intensity image. The gradient is calculated using the derivative a Gaussian filter. The method uses two thresholds, to detect strong and weak edges and includes the weak edges in the output only if they are connected to strong edges. This method is therefore less likely than the others to be fooled by noise, and more likely to detect true weak edges. The experimental results using canny edge detector are shown in figure 5. The edges ear have definitely been detected in extracted region which has higher probability to contain ear. After edge detection, connected component labeling is applied, so extra unconnected small edges are removed. We have used an 8-connected neighborhood to label a pixel. The result connected component labeling is shown in figure 6. From this result where maximum connected edges are found a rectangle is drawn around those edges. This rectangle shows detected ear. Figure 7 shows the final result this method. Fig 5 : Edge detection 51

3 2.1.5 Ear Verification Using SVM Tool A Support Vector Machine (SVM) [10, 11] is one such method that can perform pattern recognition; its use, though is not limited to this one application. While most classifiers work on Fig 6 : Connected Component Labeling Fig 7 : Final output designing rules that will place decision boundaries between data belonging to different classes, SVM goes a step ahead and designs Support Vectors such that the data belonging to different classes is now separated by a region rather than just a hyper plane. Thus, the distinction between classes is made more obvious, in an intuitive sense. In this work, we validate the detected ear using SVM classification tool. The libsvm-2.91 [12] has been used in this verification task. Here we verify using SVM tool that weather our detection is correct or not. The choice the appropriate Kernel parameter SVM for a specific application is problem dependent and ten a difficult task. Generally it is expected that if the data is known to be not linearly separable, a non-linear kernel performs better than the one based on a linear kernel. The model is created by features which are generated using detection results. The different kernel like Linear, Polynomial, Radial Basis Function, Sigmoid and accordingly parameters i.e. degree, gamma, cost co-efficient are adjusted manually to improve classification (verification) accuracy. 2.2 Template Matching Side image Skin Segmentation Ear localizatio n using ear template Get nose point Verification the detected ear Extract sector containing ear Ear position A block diagram for this method is shown in figure 8.The template matching approach presented here is an enhancement the approach presented in [4, 13]. Our presented ear detection technique involves three steps namely detection probable area ear, as described in the edge detection based method, ear localization and ear verification. In this method, instead moving the template over the entire image we first detect the area having maximum probability the ear. This step is added to the previously presented method [4], in order to reduce the time for ear detection. In our first step, skin segmentation is performed to eliminate all non-skin pixels from the image. Then nose point is identified and using distance estimation between nose tip and ear the probable area ear is found. Second step employs an f-line created template to detect ears. Third step is about to verify the detected ears. In our presented approach, the detected ear is verified using a SVM machine learning tool in addition to Euclidean distance. Here the first step that is to find probable area ear is described in the edge detection based method. The second and third steps are described as follows Ear Localization The three steps involved in ear localization process [4] are discussed in following subsections Template Creation For any template based approach, it is very much necessary to obtain a template which is a good representative the data. In this technique, ear template is created by averaging the intensities a set ear images. Human ear shape can broadly be categorized into four classes: triangular, round, oval, and rectangular [4]. For creation ear template, all types the ear are considered to obtain a good representative template. Intuitively, it seemed reasonable to us that the best template to use would be one derived by somehow averaging the some ear images the dataset that would likely be in the testing images. We would like to find a good subset the ears found in the training images that are clear, straight, and representative typical lighting/environmental conditions. It is also important that these images be properly aligned and scaled with respect to one another. To this end, considerable time was spent for manually segmenting, selecting, and aligning ear images. At the end 17 ear images were chosen. These cropped images were first converted into grayscale and then the average was found which gives final template. The ear template T is formally defined as, where N is the number ear images used for ear template creation and E k is the K th ear image. E k (i, j) and T(i, j) represent the pixel values the (i, j) th pixel E k and T respectively. Thus, our final template for ear detection is a result averaging together the 17 ear images. The actual template used in the matched filtering is size pixels. The template generated and used in the experimentation is shown in figure 9. (1) Fig 8: Block diagram the template matching based method Fig 9 : Created template image 52

4 Resizing Template It is observed that the size the ear is proportional to the size the side image. This observation is used for resizing the ear template in the proposed technique. To handle the detection ears various sizes, ear template need to be resized to make it appropriate for the detection ear in the image. By keeping the aspect ratio the ear template same, it is resized to the width obtained in above Equation 2. Where wi f and wr f be the widths the input image and the reference image respectively and wi e and wr e are the widths input ear image and the reference ear image respectively Localization To search an ear in the image I, ear template T is moved over the probable area ear in the image and normalized cross correlation coefficient (NCC) [4] is computed at every pixel. NCC at point (x, y) is defined in Equation (3) as, where sum is performed over u,v under the window containing T positioned at (x,y). Ī x,y and are the average brightness values the portion the target image under the template and template image respectively. Values NCC lie between -1.0 and 1.0. Where it is found maximum, a rectangle is drawn around it to show detected ear. Value NCC closer to 1 indicates a better match Ear verification To determine whether a detected ear is a true ear or not two methods are used namely verification using Euclidian distance [4] and verification using SVM tool [10, 11, 12] Ear Verification Using Euclidian Distance Here to determine whether a detected ear is a true ear or not, shape based ear verification is performed. To measure the similarity, Euclidian distance between the two sets (one for template and another for detected ear) mean is used, which is estimated as follows: (2) (3) Distance= (4) where M T and M E are mean ear template and detected ear respectively. Edge images the ear template and the detected ear image are obtained using canny edge detector and the similarity distance between them is calculated using Equation (4). If the value distance is less than a pre estimated threshold, detection is accepted otherwise it is rejected Ear Verification Using SVM Tool Here libsvm-2.91 tool [12] is used for ear verification. The choice the appropriate Kernel for a specific application is again problem dependent and ten a difficult task. The different kernel like Linear, Polynomial, Radial Basis Function, Sigmoid and accordingly parameters i.e. degree, gamma, cost co-efficient are adjusted to improve the verification accuracy. 3. EXPERIMENTAL RESULTS 3.1 Data Acquisition In this work CVL (CVL is library for image and data processing using graphics processing units (GPUs)) dataset [14] is used, which contains total 114 persons with 7 images each. Resolution each image is 640 x 480. All the Images are in JPEG format captured by Sony Digital Mavica under uniform illumination, and with projection screen in background. Age most the s is between years approximately. Although the method is tested on males certain age groups, it can also be applied with persons other age groups. An another dataset is produced by us, having images 30 side s with dynamic lighting condition with screen resolution Ear Results Edge Based This method is tested on 130 side images. In results canny edge detector, there are some extra edges which are not ear are also detected but the whole ear is detected with continuous edges. After edge detection threshold is applied for connected component labeling to remove unconnected small edges and a rectangular box is drawn around the place where maximum connected edges are found. If the detected region contains part ear, it is considered as a positive detection; otherwise it is a false detection. Figure 10 shows examples positive detection. This method has failed in some cases, especially for the images where is oriented with some angle because in that image nose tip detection is not correct. For some people the distance between nose tip and ear is differ than the generally estimated distance so there are chances for false detections or partial true detections. Figure 11 shows examples false detection. In some figures the nose tip detection is wrong so there is false detection, while others are having partial ear detection because variation distance between nose tip and ear compared to estimated distance. The accuracy is calculated as: (genuine localization/total sample) 100. The accuracy the presented method on the above mentioned database is obtained to be 83%. The average time to detect an ear from a side image using this method is 1.35 seconds with Matlab environment. It should be noted that this time is much less than the traditional template based approach [4] in which a template ears is moved over the entire image Template Matching To create ear template, a set ear images 17 people is considered. NCC is used to localize the ear. Points having maximum NCC values are declared as the detected ear. This experiment is performed on 100 images CVL dataset and 30 images general dataset. After experimentation, if the detected region contains part ear, it is considered to be a positive detection; otherwise it is a false detection. Figure 12 shows examples some the positive detections. The proposed technique is also able to detect ear in presence little occlusion due to hair. The fourth image fourth row figure 12 shows such example. Figure 13 shows examples some 53

5 3.3.2 Template Matching Here also libsvm-2.91 tool is used for ear verification. The template is size and so the detected ear is also the Fig 10 : Positive detection Figure 12 : Positive detection Fig 11 : False detection the false detections. Localization method has failed in some cases, especially for the images which are poor quality or heavily occluded due to hair (Fourth image second row in figure 13) or is oriented with some angle (Third image first row and third image second row in figure 13). Accuracy the localization is defined by (genuine localization/total sample) 100. It is found to be 78% for the CVL dataset and 70% for general dataset. The average time to detect an ear from a side image is approx seconds with Matlab environment. 3.3 Ear Verification Results Edge Based In this work, for ear verification libsvm-2.91 tool is used. The rectangle drawn to show detection is size So, 1600 pixel intensity values are in one feature vector. The feature vectors are created from the pixels the detected ear and model is created from these feature vectors. The different parameters SVM are adjusted for accuracy measurement. The maximum accuracy is found 75.86% using polynomial, RBF and sigmoid kernel with certain parameter settings. Using polynomial kernel with degree 1 and cost 100 the found accuracy is %. The same accuracy is found using RBF kernel with Gamma and cost and using sigmoid kernel with co efficient 1, Gamma and cost 1. Using Linear kernel the found accuracy is very less around %. Figure 13 : False detection same size. Therefore, 1600 pixel intensity values are in one feature vector. The different parameters are set for accuracy measurement. The maximum accuracy is found to be 93.33%. The maximum accuracy was obtained with RBF kernel using parameter values C and gamma to be 60 and respectively. However, in our experimentation few other combinations were also found which gave the same accuracy. 4. COMPARISON OF METHODS For CVL dataset, using edge detection based algorithm, the accuracy is around 83% and using template matching algorithm it is 78%. Some false positive detections are there. The comparison between the results for CVL dataset these two methods is shown in Table 1. Using edge detection based algorithm, satisfactory results are obtained for different environments. There are also a few false detections due to distance variation between nose tip and ear for different persons. In template matching algorithm, to detect ear in the general environment a new template has to be generated which after being used in the matching algorithm gives adequate results. The comparison between results both the methods for general dataset is shown in Table 2. 54

6 Name Algorithm Edge Based Template Matching Edge Based Template Matching Table 1. Comparison for the CVL Database No Images Detected Ears Accuracy (%) False Positive Rate Time In Seconds Table 2. Comparison for the general Database No Images Detected Ears Name Algorithm Accuracy (%) False Positive Rate Time In Seconds CONCLUSION From the results we can conclude that for edge detection based method the nose tip detection is very important because in this method ear detection is based on distance estimation between nose tip and ear. So if skin segmentation is not done properly then there are chances for wrong nose tip detection and hence causes false detection. For some people the distance between nose tip and ear differs than generally estimated distance, so there may be cases for partial ear detection. Template matching algorithm does not have any effects on the detection ears from different environments. The constraint for getting very good result is that the template has to be recreated for different datasets otherwise it degrades the performance detection. Instead moving template over the entire image, if it is moved over the region which has higher probability to contain ear then it takes less detection time. This method fails if ears are heavily occluded due to hair. In our results, both methods fail, if side is oriented with some angle. 6. REFERENCES [1] B V Srinivasan Ear Extraction From the Image a Human Face. University Maryland, College Park. [2] H. Chen and B. Bhanu, Hman Ear From 3D Side Face Range Images. 3D Imaging for Safety and Security, vol.35, Springer-2007, pp [3] S. M. S. Islam, M. Bennamoun and R. Davies, Fast and Fully Automatic Ear Using Cascaded AdaBoost Proc. IEEE Workshop on Application Computer, [4] S. Prakash, U. Jayaraman and P. Gupta, A Skin-Color and Template Based Technique For Automatic Ear. Proc. Int l Conf. Advances in Pattern Recognition, ICAPR' 09, Feruary 2009,pp [5] K. Joshi, and N. Chauhan, An Ear and Support Vector Machine based Approach for Human Ear and Varification, Int. Conf. Intellignet Systems and Data Processing (ICISD-2011), G. H. Patel College Engineering and Technology, Gujarat, India, January, [6] SL Phung, A. Bouzerdoum, D.Chai, Skin Segmentation Using Color Pixel Classification: Analysis and Comparison, IEEE trans. on Pattern Analysis and Machine Intelligence,vol.27,no.1, January 2005, pp [7] V. Vezhnevets, V. Sazonov, A. Andreeva: A Survey on Pixel- Based Skin Color Techniques International Journal Expert Systems With Applications, vol.3, April 2009, pp [8] J. D. Foley, A. Dam, S. K. Feiner, J. F.Hughes Computer Graphics Principles and Practice, Pearson Education, [9] R. C. Gonzalez and R. E. Woods, Digital Image Processing, Pearson Education, [10] N. C. Chauhan, Y. K. Roy, Arun Kumar, A. Mittal, and M. V. Kartikeyan SVM-PSO Based Modeling and Optimization Microwave Components, Frequenz 62,2008. [11] J. A. Suykens, J. Vandewalle, and B. D. Moor, Optimal Control by Least Squares Support Vector Machines, International Journal Neural Networks,vol.14, no.1, Jan,2001, pp [12] C. C. Chang and C. J. Lin, LIBSVM A Library for Support Vector Machines, [13] K. Joshi, and N. Chauhan, A Template Matching and Support Vector Machine based Approach for Human Ear and Varification, Int. Conf. Information, Signals, and Communications (ICISC-2011), A. D. Patel Institute Technology, Gujarat, India, 5-6 February,